---

name: ddpm_pa_recovery_harness description: Run a bounded DDPM-PA proxy optimizer step and emit mechanism-faithful recovery traces and metrics.

DDPM-PA Recovery Harness

Use this skill when full DDPM-PA few-shot image generation is blocked by missing checkpoints, datasets, or image-quality evaluators, but soft-mode recovery permits a declared mechanism-faithful proxy. The skill combines the adaptation protocol, pairwise image preservation, high-frequency preservation, high-frequency MSE, and an optimizer update into an auditable reduced experiment.

Do not use this skill to claim full FID or Intra-LPIPS reproduction unless a real pretrained DDPM, target images, sampling loop, and evaluators ran. In proxy mode it validates the adaptation mechanism, not the paper's final image metric.

Inputs

• Module-plan recovery target metadata.
• Runtime handoff explaining full-runtime blockers and soft-mode permission.
• Generated skill script roots for protocol, pairwise loss, and Haar high-frequency loss.
• Lambda weights consistent with the paper: lambda2 and lambda3 in [0.1, 1.0], lambda4 in [0.01, 0.08] unless an ablation intentionally changes them.

Outputs

• training_trace.json with loss values, component losses, params_before, params_after, and optimizer evidence.
• Mechanism checks for shared noising, x0 reconstruction, image pairwise KL, high-frequency pairwise KL, high-frequency MSE, and parameter update.
• A recovery metric such as combined_loss_delta generated by an executable command.

Workflow

1. Read the module plan and runtime handoff before selecting proxy mode.
2. Construct a tiny deterministic target batch and scalar trainable adapted-denoiser parameters.
3. Use generated module scripts to compute noised/reconstructed samples, pairwise KL, and high-frequency losses.
4. Combine loss components using paper-consistent lambda weights.
5. Estimate a deterministic finite-difference gradient and update adapted parameters.
6. Save traces and mechanism checks; never hand-write final metrics outside the command.

Validation

Run:

python scripts/proxy_step.py --output /tmp/ddpm_pa_proxy.json
python -m pytest tests

The tests assert finite loss, changed parameters, complete mechanism booleans, and target-compatible output fields.

Limitations

The default script is intentionally tiny and standard-library only. It is a proxy for the DDPM-PA loss geometry and optimizer execution, not an image generation benchmark.

Cycle 1 refinement note

A zero-pairwise ablation must be recorded as an ablation rather than a valid DDPM-PA mechanism run, because it removes both source-referenced preservation losses even if the base proxy loss still decreases.